Computer-aided diagnosis of subtle signs of breast cancer: Architectural distortion in prior mammograms

Transcription

1 Computer-aided diagnosis of subtle signs of breast cancer: Architectural distortion in prior mammograms Rangaraj M. Rangayyan Department of Electrical and Computer Engineering University of Calgary, Calgary, Alberta, CANADA

2 Mammography Signs of Breast Cancer: Masses Calcifications Bilateral asymmetry Architectural distortion (often missed) 2

3 Masses Breast cancer causes a desmoplastic reaction in breast tissue A mass is observed as a bright, hyperdense object 3

4 Calcification Deposits of calcium in breast tissue 4

5 Bilateral asymmetry Differences in the overall density distribution in the two breasts 5

6 Computer-aided diagnosis Increased number of cancers detected Increased early-stage malignancies detected Increased recall rate Missed cases of architectural distortion 6

7 Architectural distortion Third most common mammographic sign of nonpalpable breast cancer The normal architecture of the breast is distorted No definite mass visible Spiculations radiating from a point Focal retraction or distortion at the edge of the parenchyma 7

8 Architectural distortion spiculated focal retraction incipient mass 8

9 Normal vs architectural distortion 9

10 Normal vs architectural distortion 10

11 Initial algorithm for detection of architectural distortion 1. Extract the orientation field 2. Filter and downsample the orientation field 3. Analyze orientation field using phase portraits 4. Postprocess the phase portrait maps 5. Detect sites of architectural distortion 11

12 12 Gabor filter ( ) ( ) fx y x y x y x y x π σ σ σ πσ 2 cos 2 1 exp 2 1, g = Design parameters = = = = y x y x l f x y x cosθ sin θ sin θ cosθ ; 2ln 2 2 ; 1 σ σ τ σ τ Gabor parameters line thickness τ elongation l orientation θ

13 Design of Gabor filters l = l 0 τ = τ 0 θ = θ 0 l > l 0 τ = τ 0 θ = θ 0 l = l 0 τ > τ 0 θ = θ 0 l = l 0 τ = τ 0 θ > θ 0 13

14 Example of Gabor filtering Log-magnitude Inverted Y channel Magnitude response of Fourier spectrum of retinal fundus image a single Gabor filter: τ = 8, l = 2.9, θ = 45 ο 14

15 Extracting the orientation field Compute the texture orientation (angle) at each pixel Gabor filtering (line detection) 15

16 Phase portraits v v v ( ) x x, y = = A + b y x y node saddle spiral 16

17 Texture analysis using phase portraits Fit phase portrait model to the analysis window Nonlinear least squares optimization A = b =

18 Texture analysis using phase portraits Cast a vote at the fixed point = A -1 b in the corresponding phase portrait map Orientation field Node Saddle Spiral real eigenvalues of same sign 18

19 Detection of architectural distortion 19

20 Initial results of detection Test dataset: 19 mammograms with architectural distortion (MIAS database) Sensitivity: 84% 18 false positives per image! 20

21 Reduction of false positives 21

22 Rejection of confounding structures Confounding structures include Edges of vessels Intersections of vessels Edge of the pectoral muscle Edge of the fibroglandular disk Curvilinear Structures 22

23 Nonmaximal suppression ROI with a vessel Gabor magnitude output Output of nonmaximal suppression (NMS) 23

24 Rejection of confounding CLS Output of NMS CLS Retained Angle from the orientation field and direction perpendicular to the gradient vector differ by < 30º 24

25 Improved detection of sites of architectural distortion Node map (without CLS analysis) Node map (with CLS analysis) 25

26 Free-response ROC analysis 100 With CLS analysis 80 sensitivity (%) Without CLS analysis False positives per image 26

27 Effect of condition number of matrix A on the orientation field Condition Number: The ratio of the largest to smallest singular value of a matrix 27

28 Results 19 cases of architectural distortion 41 normal control mammograms (MIAS) Symmetric matrix A: node and saddle only Condition number of A > 3: reject result Sensitivity: 84% at 4.5 false positives/image Sensitivity: 95% at 9.9 false positives/image 28

29 Prior mammograms Detection mammogram 1997 Prior mammogram

30 Prior mammograms Detection mammogram 1997 Prior mammogram

31 Prior mammograms Detection mammogram 1997 Prior mammogram

32 Interval cancer Breast cancer detected outside the screening program in the interval between scheduled screening sessions Diagnostic mammograms not available 32

33 Dataset 106 prior mammographic images of 56 individuals diagnosed with breast cancer (interval-cancer cases) Time interval between prior and detection (33 cases) average: 15 months, standard deviation: 7 months minimum: 1 month, maximum: 24 months 52 mammographic images of 13 normal individuals Normal control cases selected represent the penultimate screening visits at the time of preparation of the database 33

34 Interval cancer: site of architectural distortion Mammogram Gabor Magnitude 34

35 Interval cancer: site of architectural distortion Orientation field 35

36 Site of architectural distortion Mammogram Gabor magnitude Orientation field Node map 36

37 Interval cancer: potential sites of architectural distortion Node map Automatically detected ROIs

38 Examples of detected ROIs True-positive False-positive 38

39 Automatically detected ROIs Data Set Prior mammograms of 56 interval-cancer cases No. of Images No. of ROIs 128 x 128 pixels at 200 μm/pixel No. of True- Positive ROIs No. of False- Positive ROIs Penultimate mammograms of 13 normal cases Total

40 Feature extraction from ROIs Potential Sites of Architectural Distortion Phase Portrait Analysis (Node value) Fractal Analysis (Fractal Dimension) Analysis of Angular Spread of Power Statistical Analysis of Texture (Haralick) Structural Analysis of Texture (Laws) Feature Selection, Pattern Classification Classification of ROIs 40

41 Fractal and spectral analysis θ f TP ROI, s(x, y) Fourier power spectrum, S(u, v) Power spectrum in polar coordinates, S(f, θ) Radial frequency spectrum, S(f) Angular spread of power, S(θ) 41

42 Laws texture energy measures Operators of length five pixels may be generated by convolving the basic L3, E3, and S3 operators: L5 = L3 * L3 = [ ] (local average) E5 = L3 * E3 = [ ] (edges) S5 = -E3 * E3 = [ ] (spots) R5 = -S3 * S3 = [ ] (ripples) W5 = -E3 * S3 = [ ] (waves) 2D 5 5 convolution operators: L5L5 = L5 T L5 W5W5 = W5 T W5 R5R5 = R5 T R5 etc.

43 Laws texture energy Sum of the absolute values in the filtered images in a window L5L5 E5E5 S5S5 W5W5 R5R5

44 Geometrical transformation for Laws feature extraction 44

45 Analysis of angular spread: True-positive ROI Frequency domain Gabor magnitude Gabor orientation Coherence Orientation strength

46 Analysis of angular spread: False-positive ROI Frequency domain Gabor magnitude Gabor orientation Coherence Orientation strength

47 Results with selected features Classifiers AUC using the selected features with stepwise logistic regression FLDA (Leave-one-ROI-out) 0.75 Bayesian (Leave-one-ROI-out) 0.76 SLFF-NN (Single-layer feed forward: tangent-sigmoid) 0.78 SLFF-NN*(Single-layer feed forward: tangent-sigmoid) 0.78 ± 0.02 * Two-fold random subsampling, repeated 100 times 47

48 Free-response ROC Sensitivity 80% at 5.8 FP/image 90% at 8.1 FP/image using features selected with stepwise logistic regression, the Bayesian classifier, and the leave-oneimage out method 48

49 Bayesian ranking of ROIs: unsuccessful case

50 Bayesian ranking of ROIs: successful detection 50

51 Geometrical analysis of spicules and Gabor angle response Index of convergence of spicules P Q: size of the ROI θ(i, j): Gabor angle response within the range [-89, 90 ] M(i, j): Gabor magnitude response α(i, j): angle of a pixel with respect to the horizontal toward the center of ROI, in the range [-89, 90 ]

52 ICS quantifies the degree of alignment of each pixel toward the center of the ROI weighted by the Gabor magnitude response Index of convergence of spicules

53 FROC analysis Sensitivity 80% 5.3 FP/patient 90% 6.3 FP/patient

54 Expected loci of breast tissue 54

55 Landmarking of mammograms: breast boundary, pectoral muscle, nipple Second- and fifth-order polynomials fitted to parts of breast boundary 55

56 Derivation of expected loci of breast tissue: interpolation 56

57 Number of points in curve = M L i = length between two curves at the i-th point L max = max(l i ) Number of curves = N = L max +1 Distance at i-th point = L i /L max = L i /(N-1) i-th point of n-th curve: 57

58 Divergence with respect to the expected loci of breast tissue M: Gabor magnitude response ɵ: Gabor angle response ɸ: expected orientation of breast tissue L: 25 pixels at 200 μm/pixel 180 Gabor filters used over [-90, 90] degrees 58

59 Orientation field of breast tissue obtained using Gabor filters Original image Gabor magnitude Gabor angle 59

60 Divergence with respect to the expected loci of breast tissue Original image Divergence map Thresholded map 60

61 Automatically detected regions of interest ROC: AUC = 0.61 FROC: Sensitivity = 80% at 9.1 FP/patient 61

62 Combination of 86 features Geometrical features of spicules: 12 Haralick s and Laws texture features, fractal dimension: 25 Angular spread, entropy: 15 Haralick s measures with angle cooccurrence matrices: 28 Statistical measures of angular dispersion and correlation: 6 Feature selection with stepwise logistic regression Bayesian classifier with leave-one-patient-out validation: 80% sensitivity at 3.7 FP/patient 62

63 Reduction of false positives

64 Reduction of false positives 64

65 Conclusion Our methods can detect early signs of breast cancer 15 months ahead of the time of clinical diagnosis with a sensitivity of 80% with fewer than 4 false positives per patient Future work: Detection of sites of architectural distortion at higher sensitivity and lower false-positive rates Application to direct digital mammograms and breast tomosynthesis images 65

66 Thank You! Natural Sciences and Engineering Research Council (NSERC) of Canada Indian Institute of Technology Kharagpur Shastri Indo-Canadian Institute University of Calgary International Grants Committee Department of Information Technology, Government of India My collaborators and students: Dr. J.E.L. Desautels, N. Mudigonda, H. Alto, F.J. Ayres, S. Banik, S. Prajna, J. Chakraborty, Dr. S. Mukhopadhyay 66